# Nvkvm Pv

> Paravirtual NVIDIA GPU for KVM guests — run unmodified CUDA, PyTorch and Vulkan inside a VM at host parity, on a GPU the host keeps using.

- **Website:** https://github.com/reindertpelsma/nvkvm-pv
- **Pricing:** unknown
- **Categories:** Developer Tools, Design
- **Tags:** developer-tools, design, infrastructure
- **Platforms:** Windows, Linux, iOS, CLI
- **Last verified:** 2026-09-10
- **Canonical page:** https://linkrena.com/tools/nvkvm-pv

## About

Run CUDA, PyTorch, Vulkan or a graphical display inside a KVM guest, on the same GPU your host is still using.

Get the actual GPU accelerated display in the VM on your desktop - zero copy native, no PCIe round trips when not needed.

Run QEMU in a container for additional isolation and environment consistency.

This project is both intended for compute workload VMs (e.g deep learning) and for graphics VMs (e.g gaming)

Without /dev/kvm , QEMU silently falls back to software emulation: it appears to work and is unusably slow. Permissions, containers and nested virt can be found here: install guide .

Keep the 127.0.0.1: in the -p flag. The guest is ubuntu:ubuntu with NOPASSWD:ALL ; published as a bare -p 2222:2222 it is a root shell with your GPU attached, offered to your whole network.

NVIDIA_DRIVER_CAPABILITIES is not optional decoration: --gpus all alone gives the container compute,utility , and the guest then gets a compute-only driver with no GL or Vulkan — which does not fail, it silently falls back to llvmpipe.

A prebuilt tarball runs on a bare host, and bash scripts/build_qemu.sh --install-deps builds from source.

## Related tools

- [Pollen Robotics](https://linkrena.com/tools/pollen-robotics): Microduck is a 25 cm biped robot with 15 motors, a camera, LiDAR and a grasping beak.
- [Derive](https://linkrena.com/tools/derive): Review and approval for work made by AI agents.
- [Spore](https://linkrena.com/tools/spore): Run open-weight models on hardware you already own, reach them from any device, and tap a distributed network when you need more.
- [Conductai](https://linkrena.com/tools/conductai): AI agent governance for teams.
- [Splatit](https://linkrena.com/tools/splatit): An Splatoon server recreation in C++ with a focus on user friendliness and ease of hosting.
- [The Simpsons Hit and Run](https://linkrena.com/tools/the-simpsons-hit-and-run): Modern Windows version of The Simpsons: Hit & Run, based on the original leaked source code from 2003 - xaskasdf/The-Simpsons-Hit-and-Run